Senior researcher

The approach involves using limited sinogram data, obtained from electronic microscopy instruments, as
input to image-to-image regression networks that are trained to interpolate the missing data, followed
by image-to-image AI networks to denoise those results. We will use a subset of the full sinogram data
as input to an in-painting neural network that will be selected based on sinogram angles containing the
most overlap information. Reconstructed image from the predicted sinograms should match the original
reconstruction from the full dataset. This approach has been studied in the literature for less demanding
problems and will help us minimize the number of required experimental acquisitions.

We will also explore the idea of using AI image-to-image networks that take sinogram data as input and
directly output the reconstructed images, thereby minimizing the artifacts produced by iterative
reconstruction methods. We will pre-train image-to-image regression networks on full data sets
(imaging at every angle between 0 and 180). We will train networks on simulated data for which we
have ground truth reconstruction. Using these pre-trained networks, we will re-train using limited
datasets and the same ground truth labels. This approach has also been seen in the literature for less
complex image reconstruction problems. We will build up our models gradually, beginning with
networks that reconstruct 2D images to perfect our methods for this type of reconstruction, and then
extend these methods to full 3D reconstructions.

Work Location: Physically at NIST (Gaithersburg, MD),Physically at University,Telework